From 8d694a037c20642f3d728800d711876be0b48956 Mon Sep 17 00:00:00 2001
From: Guy Jacob <guy.jacob@intel.com>
Date: Thu, 24 Jan 2019 17:25:30 +0200
Subject: [PATCH] Bugfix in test_quantizer
---
tests/test_quantizer.py | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/tests/test_quantizer.py b/tests/test_quantizer.py
index 49f4cd5..6865d57 100644
--- a/tests/test_quantizer.py
+++ b/tests/test_quantizer.py
@@ -317,15 +317,17 @@ def test_param_quantization(model, optimizer, qbits, bits_overrides, explicit_ex
if has_children(pre_quant_module):
continue
- num_bits = expected_qbits[name].wts
+ num_qbits = expected_qbits[name].wts
for param_name, pre_quant_param in pre_quant_module.named_parameters():
- quantizable = num_bits is not None
+ quantizable = num_qbits is not None
if param_name.endswith('bias'):
quantizable = quantizable and quantize_bias
# Bias number of bits is hard-coded to 32 for now...
if quantizable:
num_bits = 32
+ else:
+ num_bits = num_qbits
if quantizable and train_with_fp_copy:
# "param_name" and "pre_quant_param" refer to the float copy
--
GitLab